Research map / Reinforcement learning / RL from feedback
Joint Alignment and Distillation for Video Generation via Sample-Guided Distribution Matching
Jiuzhou Lin, Fei Zuo, Huan Ouyang, Junlong Wu, Dewen Fan, Boheng Zhang, Huaiqing Wang, Jia Sun, Fan Yang, Houde Liu, Kehai Chen, Min Zhang and 2 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| single-stage DM-Align (Pair) instead of raw model | Wan2.1-T2V-1.3B · VBench (ConsistID, Pair) | Average score: 78.20 -> 82.78 | Improved |
| single-stage DM-Align (Group) instead of raw model | Wan2.1-T2V-1.3B · VBench (VidProM, Group) | Average score: 77.85 -> 84.40 | Improved |
| apply DMD2 distillation after Flow-DPO (two-stage) | Flow-DPO · VBench (ConsistID, Pair) | Dynamic Degree: 67.13 -> 53.43 | Hurt |
| apply DMD2 distillation after DanceGRPO (two-stage) | DanceGRPO · VBench (VidProM, Group) | Dynamic Degree: 75.78 -> 60.85 | Hurt |
| joint DM-Align instead of two-stage Flow-DPO + DMD2 | DM-Align (Pair) · VBench (ConsistID, Pair) | Average score: 79.89 -> 82.78 | Improved |
| joint DM-Align instead of two-stage DanceGRPO + DMD2 | DM-Align (Group) · VBench (VidProM, Group) | Average score: 80.54 -> 84.40 | Improved |
| DM-Align vs DanceGRPO training step cost | DM-Align · Training efficiency | time per step (s): 423 -> 15 | Improved |
| apply DM-Align to Wan2.2-A14B MoE (high-noise expert) | Wan2.2-T2V-A14B · VBench (Wan2.2-A14B MoE) | VBench score: 78.35 -> 83.45 | Improved |
| add alignment loss with 1:1 weight to DMD (ablation) | DM-Align (Wan-1.3B) · Loss weight sensitivity (HPSv2) | HPSv2: 20.50 -> 28.89 | Improved |
| raise alignment weight to 1:3 (low DMD weight) (ablation) | DM-Align (Wan-1.3B) · Loss weight sensitivity (HPSv2) | HPSv2: 28.89 -> 12.37 | Hurt |
| linear scheduling of weights from 1:0 to 1:3 instead of 1:1 (ablation) | DM-Align (Wan-1.3B) · Loss weight sensitivity (TA) | TA: 1.65 -> 1.45 | Hurt |
| add DM-Align on CogVideoX-2B vs DMD only (ablation) | CogVideoX-2B · Loss weight sensitivity (HPSv2) | HPSv2: 18.71 -> 25.01 | Improved |
Closest papers by meaning
- Aligning Human Sense: Calibrated Distributional Reward Learning for Video Generation (ECCV 2026 · similarity 0.91)
- Threshold-Guided Optimization for Visual Generative Models (ICML 2026 · similarity 0.91)
- Improving Video Generation with Human Feedback (NeurIPS 2025 · similarity 0.90)
- VideoDPO: Omni-Preference Alignment for Video Diffusion Generation (CVPR 2025 · similarity 0.88)
- Aligning Few-Step Generative Model via Amortizing Sample-Based Variational Inference (NeurIPS 2026 · similarity 0.88)
- David and Goliath: Small One-step Model Beats Large Diffusion with Score Post-training (ICML 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
